Reverse design for mixture proportion of asphalt concrete in high-temperature regions based on the materials informatics

Abstract

With the development of economy and living quality, the requirements of high-temperature performance of asphalt mixture are constantly increasing. However, the traditional asphalt mix design method is difficult to meet these requirements due to its cumbersome and blind nature. This study aims to implement the concept of materials informatics to the design of asphalt concrete in high-temperature regions. Firstly, the data of asphalt binders and mixtures containing modulus and conventional property information were obtained from an open database, and outliers were detected and removed. Every feature was redefined based on materials informatics and selected based on prior knowledge. Six machine learning models were then employed to develop forward design models, including ridge regression, K-nearest neighbor regression, support vector machine regression, multilayer perceptron, random forests, and the stacking model, which established the mapping relationship from input to output for the reverse mix design. Finally, by leveraging the stacking model and genetic algorithm, the multi-objective optimization was conducted to achieve a trade-off between dynamic modulus and cost of the asphalt mixture. The results indicate that the feature selection based on prior knowledge was feasible. All forward design models can achieve promising prediction accuracy, among which the stacking model has the best generalization ability, with the lowest variance and bias. Lastly, the multi-objective optimization method based on materials informatics could efficiently identify the feasible ranges of key design parameters for low-cost asphalt mixtures that reach the requirements of target service performance, providing a valuable solution to dilemmas existing in the mix proportion design of asphalt concrete in the high-temperature regions.